Brute force face feature point recognition method based on large-scale classifier

Abstract

The present invention provides a brute force face feature point recognition method based on a large-scale classifier. The main content of the method comprises the steps of: regression analysis throughadoption of large-scale classification, a solution strategy of classification problems, a preprocessing process and a post-processing process. The method concretely comprises the steps of: decomposing all the possible prediction conditions into a plurality of discrete categories, wherein the step consists of clustering and probability reasoning; employing a multi-tag frame to train a large-scalemulti-class network, and employing an Softmax network layer to replace a loss network layer to cope with a decision boundary problem; and finally, employing a preprocessing process to learn a linear bounding box regressor variable, and adding a post-processing regression step to improve the efficiency. Based on the large-scale classifier, the brute force face feature point recognition method is provided to avoid the condition that the regression problem has many solutions, and is higher in robustness and more accurate in face feature point recognition rate.

Description

technical field [0001] The invention relates to the field of face feature point recognition, in particular to a brute force face feature point recognition method based on a large-scale classifier. Background technique [0002] Facial feature point recognition is a research topic about the recognition and location of facial feature points (such as eyes, nose, mouth, etc.). Facial feature point recognition technology is widely used, such as face alignment, face deformation, face recognition, facial expression analysis, human-computer interaction, etc. For example, you can use the face feature point recognition technology to obtain the feature points of a face photo, and then drive it to deform through the feature points to find the face that is closest to it but more in line with human aesthetics, so as to beautify the face Effect. In addition, accurate face feature point recognition and positioning can greatly improve the accuracy of face recognition, which is of great sign...

AI Technical Summary

Problems solved by technology

[0004] Aiming at the shortcomings of existing methods that may have multiple solutions, the method is not robust enough and not accurate enough, the purpose of the present invention is to provide a brute force face feature point recognition method based on a large-scale classifier. First, all possible predictions The situation is decomposed into many discrete categories. This process includes two steps of clustering and probabilistic reasoning; then use the multi-label framework to train a large-scale multi-category network, and use a Softmax network layer instead of the loss network layer to cope with the decision boundary. problem; finally, use the preprocessing process to learn a linear bounding box regressor, and add a postprocessing regression step to improve accuracy

Images